Are we outsourcing our judgement to AI?

Your phone suggests the reply to a text before you've decided what you think. Your inbox drafts the email and you just press send. You ask ChatGPT what to say to your landlord, and it tells you.

Not one of those tools stopped you deciding. But it can feel as if the decision has already been made, and the only thing left is to click accept.

Whether AI is useful is not in doubt. A question worth asking is whether you still own the judgement once you accept the output.

The slip is easiest to see with recommendations. Faced with something unfamiliar, it is natural to ask the tool what it would do. The answer arrives quickly, sounds reasonable and saves you working it out. That is a good starting point. The risk comes when you accept the recommendation without working out whether its reasons make sense in your situation. That is the moment judgement starts to leave the room.

Handing thinking to tools is not new

Using something outside your own head to reduce the mental work is called cognitive offloading. Setting a reminder counts. So does writing a shopping list. Neither hands over any judgement. The question is which parts of the thinking we are handing over. Risko & Gilbert, 2016.

Offloading did not begin with ChatGPT.

Calendars hold appointments. Satnav picks the route. Spreadsheets do the arithmetic. Spam filters sort the inbox. Support software routes tickets.

Those tools carry assumptions too. Satnav decides that the fastest route is the one you want. A spam filter has a view on what you should see. But they work inside a narrow job you understand well enough to overrule them. What has changed is scope. With conversational AI you can ask for advice on almost anything, including things you do not know enough about to judge the answer.

The newer tools reach further. An ad platform decides which audiences see your campaign and how the budget is split between them. A helpdesk suggests not just a reply but the tone to take with a frustrated customer. An assistant reads a forty-page report and tells you the three things that matter.

Each of those is a judgement. Which customer deserves more care. What counts as important. The tools deliver those judgements in language that sounds like careful thought, and the risk is treating the output as the judgement itself, rather than as material for one.

Two questions to ask before you hand something over

Researchers at MIT's Center for Information Systems Research built a framework around two questions. How much room is there for interpretation? And how bad would a mistake be? MIT Sloan summary · Full briefing

Together they give four types of decision: routine, consequential, exploratory and strategic. Each has a different role for AI.

Routine decisions (low ambiguity, low risk): tagging enquiries, sorting receipts. The rule is clear and a mistake costs a minute. Automate them.

Consequential decisions (low ambiguity, high risk): the task looks routine but the consequences are not. A £5 refund and a £5,000 refund are the same task to the tool. Keep a person in the loop wherever a wrong answer costs money or trust.

Exploratory decisions (high ambiguity, low risk): drafting, brainstorming, prototyping, anything where a wrong turn is cheap. Let the tool widen your options. A poor suggestion only costs the time it takes to read it.

Strategic decisions (high ambiguity, high risk): whether to drop a service line, how to handle a complaint that could go public. People frame the problem and own the outcome. AI gathers the options.

What concerns me is how easily a consequential decision can look like a routine one. A customer reply or a price quote feels like the same kind of task as sorting receipts. It is not. Meanwhile the exploratory work is where the tool can widen your thinking at almost no cost.

How do we keep control of our judgement?

Frameworks help you decide what to automate. Habits decide whether you still understand the call.

The test is simple. If you cannot explain your reasoning for a decision without the tool open in front of you, the work is not finished.

When a recommendation looks polished, ask what it is based on, then check the claim yourself. If AI summarises a report, open the report and confirm it says what the summary claims. A model can state a wrong fact as confidently as a right one, and the polish makes the mistake harder to spot.

Then check the trade-offs. A recommendation can be accurate and still prioritise something you do not care about. Ask what it assumed you wanted, and whether that is true.

None of this means refusing suggestions. It means refusing to accept one you do not understand. See our AI content service for how we keep people involved in reviewing marketing content.

Own the framing, the trade-offs and the call

We have always outsourced memory and calculation. Handing over a first draft or a recommendation can sit in that same tradition, as long as you still decide what to do with it.

Judgement is only outsourced when someone stops framing the problem, stops weighing the trade-offs, and treats the model's output as the decision.

Keep the tool. Keep the understanding.

For more practical marketing perspectives, read our Go Further newsletter.

Want to discuss AI in your business? Talk to Lunr Launch about your marketing.